Aiding diagnosis of rare diseases from photographs using machine learning

The objective of this thesis is to further research in the field of computational syndrome diagnosis from ordinary photographs of patient faces. As well as providing insight into the phenotypic scope of specific genetic disorders, the development of such tools may eventually help to shorten the process for patients to receive a clinical diagnosis. The success of computer vision algorithms in this domain crucially depends on sourcing large labelled data sets of patient syndrome faces. In this thesis, we address this issue by creating a download pipeline to obtain facial images of genetic disorder patients from the published biomedical literature. The development of this pipeline has led to the creation of a novel biomedical figure classifier; and a syndrome annotation tool, which enables rapid diagnostic labelling of retrieved patient images. Critical to modelling the craniofacial manifestations associated with genetic disorders is the need to separate phenotypic signals caused by the genetic disorder from the typical variation found among inherited facial features. Unfortunately, existing kinship data sets display several biases which undermine the task of facial kinship modelling. To demonstrate this, we build a classifier which can determine whether two facial images are cropped from the same original photo. Despite this classifier having no understanding of kinship, it is able to achieve comparable performance to many previously published methods on several existing kinship data sets. To counter this, we construct two novel kinship data sets collected from the Geni ancestry website, being careful to avoid several common biases found in other kinship data sets. Finally, we attempt to benchmark the kinship signal in our newly produced data sets using feature aggregation methods. The data sets and techniques discussed in this thesis are currently being used to produce the next generation of diagnostic syndrome face models.